GNBAN Improves Long-Horizon Forecasting for Retail Demand

GNBAN, or Graph Neural Basis Attention Network, is a novel AI architecture designed to tackle the complex challenge of long-horizon demand forecasting, particularly for large retail hierarchies. Traditional forecasting methods often struggle with the scale and interconnectedness of tens of thousands of correlated time series across products, stores, and regions. GNBAN addresses these limitations by representing retail data as a heterogeneous graph, allowing a single model to serve an entire catalog and capture intricate cross-entity dependencies. A key innovation of GNBAN is its interpretable basis-decomposition head, which breaks down each forecast into distinct trend, seasonal, and generic components. This enhances transparency, a crucial factor for trust in forecasting systems. The model employs a unique per-basis attention mechanism, where each basis function independently retrieves information from an entity's historical neighborhood, enabling specialization for different temporal patterns. Evaluated on large-scale benchmarks like M5 Walmart and Favorita Grocery Sales, GNBAN demonstrated a significant improvement of 4-5% in volume-weighted WRMSSE over comparable graph baselines, while also providing clear insights into demand drivers without requiring post-hoc explanations.